Covid-19: Scoping the field

1 Data controls

Quickly checking if our data set: gwmp is integrated properly: 0, 0, 0, 0, 0, 0, 78059. Looks fine to me. We should perform later on more check ups. Like:

1. Are the number of cases (taken from Oxford’s dataset which in turn is taken from…?) in agreement with other sources.
2. Some more random weather checks.
3. …

2 Introduction

\(\color{blue}{\text{ When you see text in blue, this is me (Orestis) commenting. You can comment in green.}}\) \(\color{blue}{\text{ Oh, and let's please find a better way of commenting. This absolutely sucks.}}\)

The ultimate measure of a man is not where he stands in moments of comfort and convenience, but where he stands at times of challenge and controversy.

— Martin Luther King

Our goal in this project is to construct a new behavioural measure of ‘compliance’. Unlike most other behavioural approaches that elicit such measures with experimental procedures, our approach is to derive this measure from data capturing real-life responses. We treat Covid-19 as an exogenous shock that allows us to track real-world behaviour, namely, responsiveness to Policy (but also, to each country’s respective urgency). The high-stakes (life or death) of the situation and the attention it has received world-wide make this index all the more meaningful. Identifying this behavioural measure and its cross-regional heterogeneity has potentially immediate policy implications. At the time when these lines are written: 05 October, 2020 , the world is experiencing a resurgence of total cases infected. A combination of factors such as colder and dumper weather making outdoors gatherings harder and re-opening of schools is likely to reinforce this trend.

\(\color{blue}{\text{ We need to a) extend the observation period to September}}\)

\(\color{blue}{\text{ b) cross-validate the total cases data (and total new cases that I constructed from first differences) with another data set. Currently I took those from Oxford.}}\)

Should governments initiate a second wave of restrictive policies, it is worth asking the question: how effective will their policies be? We conjecture that the answer to this question depends by and large on citizens’ compliance to these governmental mandates.

Recent anlyses have suggested that there are benefits to be harnessed by using behavioural measures elicited cross-culturally in the past. For example (cite the LSE article: [https://blogs.lse.ac.uk/businessreview/2020/04/22/long-read-cultural-evolution-covid-19-and-preparing-for-whats-next/]) :

Figure 2.1: Collectivism by Total Death

\(\color{blue}{\text{ This is a replication using our data. We want to divide with population to replicate more appropriately. }}\) \(\color{blue}{\text{ But we are currently missing information on population on a lot of countries.}}\)

Figure 2.2: Collectivism by Total Cases

Maybe mention that although collectivism correlates highly with GDP, there is still some unexplained variation.

3 Key variables

Just as every empirical scientific endeavor, the researcher faces the challenge of transitioning from the latent to the observable to the convincing. In our case, the latent variable of interest is that of compliance to policies that aim to reduce the virus’ spread. We construct this measure by measuring the impact of governmental policies on people’s behaviour.

Thanks to the datasets of Oxford and Coronanet, we are able to use an integrated measure of stringency, defined at a country level. Thanks to Google’s initiative to publicize anonymized data on mobility, we can track changes in movement.

3.1 Stringency Index

There are two major initiatives aiming to quantify the stringency of policies implemented by governments in order to contain the spread of the disease, the CoronaNet research project, and the Oxford COVID-19 Government Response Tracker (OxCGRT). The following two subsections aim at exploring the data offered by those projects and assess their suitability for our aims.

3.1.1 CoronaNet research project

The CoronaNet project produces a dataset that reports individual policy actions by Governments around the world, relying on a network of volunteers around the world who collect and submit interventions based mainly on media reports. The authors also derive an aggregate measure of policy activity, which uses “a statistical model that takes into account which policies seem to be more costly/difficult to implement than other policies” which is described in more detail in their Nature Human Behaviour publication.

In its raw form the dataset has one entry per policy action - this could be several per day for any given country, or none. To try and get a sense of the dataset, let us first have a look at the types of entries that dominate it.

Most entries describe intervenrions related to health resources, followed by external border restrictions, closure of schools, restriction of non-essential business, and quarantine/lockdown.

Each of these types of events is divided into sub-types. For instance, ‘Health resources’ breaks down as follows:

Of particular interest to our endeavour is the general category Lockdown/Quarantine, so let us have a closer look.

Restricting the dataset to this type of event, we are left with 1135 observations - way more entries than what we would expect given what we normally understand by ‘lockdown’.

The vast majority of these interventions are registered as ‘mandatory’, so seem to go beyond mere recommendations by government agencies or officials (which could legitimately be repeated several times). Further slicing of the data reveals that most of the interventions are at the national level, meaning that this doesn’t seem to be driven by federalist countries, where lockdowns happen at a sub-national level (e.g., US and Germany).

The temporal distribution of the Quantine/Lockdown events registered in the CoronaNet database roughly corresponds to the perception that policy activity reached its peak in late March:

For a systematic comparison of policies across countries, this type of data of course needs further processing and aggregating. Given CoronaNet’s ambition of digesting such a large number of events (and along many dimensions), this seems like a daunting task. In order to collapse the information into a single index, the authors employ Bayesian methods that essentially aim to quantify the ‘cost’ of any given type of measure, and from that ‘cost’ derive a level of policy activity. The graph below offers a glimpse at their index, aggregated by continent and week for legibility.

A few things are worth noting here. First, the range of the index - which in principle ranges from 0 to 100 - is very small. In February, when there were almost no restrictions in place anywhere, the lowest average value of the index is 44 (Africa); in April, when most countries in Europe were enforcing their strictest lockdowns, the highest average value is 49. Of course, this may be a pure scaling issue, and restricting the graph to the relevant range yields a picture that is roughly in line with increasingly strict / costly policies implemented between March and April.

Second, and perhaps more problematically, there seems to be little movement downwards. While most countries started relaxing restrictions in May and June, the index remains largely unaffected. And rather unintuituvely, the only distinct drops in the policy index happen in late February.

3.1.2 Oxford COVID-19 Government Response Tracker

OxCGRT uses a number of indicators on policy restrictions to compile a stringency index that can take values between 0 (no restrictions) and 100 (maximum restrictions).

Again, let us start by looking at the building blocks of the index. In contrast to CoronaNet, OxCGRT does not provide a database of individual evens, but instead offers almost binary time series on individual policies, which essentially flag whether a given policy is in place or not in the country. There is also rudimentary information on the ‘intensity’ of the policy (calues between 1 and 4).

The graph below plots index for a very select number of countries (chosen at random, except US and Germany), along with colorful bands that indicate whether a given policy was in place:

legend

3.2 Comparison of Policy Indices

Finally, let’s see if these datasets talk to each other at all. The plot below looks at the correlation between the two measures, based on daily values at the country level for each index.

While the two measures differ in terms of scaling and also qualitatively in recent times (after policy relaxations), the correlation is fairly high (0.5) and perfectly significant. However, the differences are clearly large enough to have an impact on any analysis, and the choice needs to be carefully considered.

3.3 Movement

Describe Google’s mobility reports.

Mobility trends seem to correspond well with Oxford’s data set.

4 Some tables and graphs

Notice how I round the digits now from the table.

5 Missing variables: Weather

We need to find a way to cite.

Going more granular: city level analysis.

Increasing the level of our data-granularity is a good thing. Even if we cluster errors at the country-level, we still get a boost on power. Moreover, it does not matter if we do not get city-level data for all variables of interest. For example, @alfaro2020social, use city-level mobility data but country-level death and total cases data. We probably should do the same.

On the map: